DeepSeek might not be as disruptive as claimed
DeepSeek might not be as disruptive as claimed, firm reportedly has 50,000 Nvidia GPUs and spent $1.6 billion on buildouts
Chinese startup DeepSeek recently took center stage in the tech world with its startlingly low usage of compute resources for its advanced AI model called R1, a model that is believed to be competitive with Open AI's o1 despite the company's claims that DeepSeek only cost $6 million and 2,048 GPUs to train. However, industry analyst firm SemiAnalysis reports that the company behind DeepSeek incurred $1.6 billion in hardware costs and has a fleet of 50,000 Nvidia Hopper GPUs, a finding that undermines the idea that DeepSeek reinvented AI training and inference with dramatically lower investments than the leaders of the AI industry.
DeepSeek operates an extensive computing infrastructure with approximately 50,000 Hopper GPUs, the report claims. This includes 10,000 H800s and 10,000 H100s, with additional purchases of H20 units, according to SemiAnalysis. These resources are distributed across multiple locations and serve purposes such as AI training, research, and financial modeling. The company's total capital investment in servers is around $1.6 billion, with an estimated $944 million spent on operating costs, according to SemiAnalysis.
DeepSeek took the attention of the AI world by storm when it disclosed the minuscule hardware requirements of its DeepSeek-V3 Mixture-of-Experts (MoE) AI model that are vastly lower when compared to those of U.S.-based models. Then DeepSeek shook the high-tech world with an Open AI-competitive R1 AI model. However, the reputable market intelligence company SemiAnalysis revealed its findings that indicate the company has some $1.6 billion worth of hardware investments.
DeepSeek originates from High-Flyer, a Chinese hedge fund that adopted AI early and heavily invested in GPUs. In 2023, High-Flyer launched DeepSeek as a separate venture solely focused on AI. Unlike many competitors, DeepSeek remains self-funded, giving it flexibility and speed in decision-making. Despite claims that it is a minor offshoot, the company has invested over $500 million into its technology, according to SemiAnalysis.
A major differentiator for DeepSeek is its ability to run its own data centers, unlike most other AI startups that rely on external cloud providers. This independence allows for full control over experiments and AI model optimizations. In addition, it enables rapid iteration without external bottlenecks, making DeepSeek highly efficient compared to traditional players in the industry.
Then there is something that one would not expect from a Chinese company: talent acquisition from mainland China, with no poaching from Taiwan or the U.S. DeepSeek exclusively hires from within China, focusing on skills and problem-solving abilities rather than formal credentials, according to SemiAnalysis. Recruitment efforts target institutions like Peking University and Zhejiang University, offering highly competitive salaries. According to the research, some AI researchers at DeepSeek earn over $1.3 million, exceeding compensation at other leading Chinese AI firms such as Moonshot.
Due to the talent inflow, DeepSeek has pioneered innovations like Multi-Head Latent Attention (MLA), which required months of development and substantial GPU usage, SemiAnalysis reports. DeepSeek emphasizes efficiency and algorithmic improvements over brute-force scaling, reshaping expectations around AI model development. This approach has, for many reasons, led some to believe that rapid advancements may reduce the demand for high-end GPUs, impacting companies like Nvidia.
What's Your Reaction?
